A Multilevel Approach to Geography of Innovation

Abstract

Srholec M. A multilevel approach to geography of innovation, Regional Studies. The aim of this paper is to demonstrate how research on geography of innovation can benefit from multilevel modelling. Using micro-data from the third Community Innovation Survey in the Czech Republic, the hypothesis that regional framework conditions determine the innovative performance of firms and that this effect differs for different kinds of firms is quantitatively assessed. The results indicate that the quality of the regional innovation system directly influences the likelihood of a firm to innovate and that this effect decreases with the size of the firm. Also, broader social characteristics of the region are relevant explanatory factors of innovation.

Srholec M. Une approche à la géographie de l'innovation à plusieurs niveaux, Regional Studies. Cet article cherche à montrer comment la recherche sur la géographie de l'innovation peut profiter de la modélisation à plusieurs niveaux. Employant des données microéconomiques provenant de la troisième Enquête sur l'innovation communautaire menée dans la République tchèque, on évalue quantitativement l'hypothèse suivante: les conditions préalables du cadre régional déterminent la performance des entreprises pour ce qui est de l'innovation et cet effet varie suivant le statut de l'entreprise. Les résultats laissent voir que la qualité du système d'innovation régional influe directement sur la possibilité que les entreprises aillent innover et que cet effet diminue suivant la taille de l'entreprise. En plus, des caractéristiques sociales régionales plus générales s'avèrent des facteurs explicatifs de l'innovation.

Innovation Géographie Système d'innovation régional Modélisation à plusieurs niveaux République tchèque

Srholec M. Ein mehrschichtiger Ansatz für die Geografie der Innovation, Regional Studies. Mit diesem Beitrag soll nachgewiesen werden, wie die Forschung über die Geografie der Innovation von mehrschichtigen Modellen profitieren kann. Anhand von Mikrodaten aus der dritten Innovationserhebung der Gemeinschaft in der Tschechischen Republik nehmen wir eine quantitative Bewertung der Hypothese vor, dass die innovative Leistung von Firmen von den regionalen Rahmenbedingungen abhängt und dass diese Auswirkung für verschiedene Arten von Firmen unterschiedlich ausfällt. Aus den Ergebnissen geht hervor, dass die Qualität des regionalen Innovationssystems einen direkten Einfluss auf die Wahrscheinlichkeit ausübt, dass die Firmen Innovationen hervorbringen, und sich dass diese Wirkung mit der Größe der Firma abschwächt. Ebenso erwiesen sich generellere gesellschaftliche Merkmale der Region als relevante Faktoren zur Erklärung von Innovation.

Innovation Geografie Regionales Innovationssystem Mehrschichtige Modelle Tschechische Republik

Srholec M. Un enfoque de varios niveles para la geografía de la innovación, Regional Studies. La finalidad de este artículo es demostrar cómo pueden beneficiarse los estudios sobre la geografía de la innovación con el modelo de varios niveles. Usando micro datos del tercer Estudio de Innovación Comunitario en la República Checa, evaluamos cuantitativamente la hipótesis de que las condiciones de estructura regional determinan el rendimiento de innovación de las empresas y que este efecto es distinto para diferentes tipos de empresas. Los resultados indican que la calidad del sistema de innovación regional influye directamente en la probabilidad de innovar de las empresas y que este efecto disminuye con el tamaño de la empresa. Asimismo se demuestra que las características sociales más generales de la región son factores explicativos relevantes de la innovación.

Innovación Geografía Sistema de innovación regional Modelos de varios niveles República Checa

Keywords:

• Innovation
• Geography
• Regional innovation system
• Multilevel modelling
• Czech Republic

JEL classifications:

• D21
• O30
• R15

Acknowledgements

The author is grateful to the Czech Statistical Office for providing access to the CIS micro-data. Special thanks go to Martin Mana and Josef Plandor for supporting access to the data set. Financial support from the post-doctoral project on ‘Innovation, Fragmentation and Economic Development’ funded by the University of Oslo is gratefully acknowledged. An earlier version of this paper was presented in the seminar series on ‘Understanding Innovation’, Centre for Advanced Study (CAS), Norwegian Academy of Science and Letters, Oslo, Norway, October 2007; the 1st Dynamics of Institutions & Markets in Europe (DIME) Scientific Conference on ‘Knowledge in Space and Time: Economic and Policy Implications of the Knowledge-based Economy’, Bureau d'Économie Théorique et Appliquée (BETA), University Louis Pasteur, Strasbourg, France, April 2008; and the DIME International Workshop on ‘Reconsidering the Regional Knowledge Economy: Theoretical, Empirical and Policy Insights from Diverse Research Approaches’, The Centre for Knowledge, Innovation, Technology and Enterprise (KITE), Newcastle University, Newcastle upon Tyne, UK, September 2008. The author thanks Björn Terje Asheim, Tom Broekel, Cristina Chaminade, Jan Fagerberg, Sjur Kasa, Ben Martin, David Mowery, Allen Scott, other participants at these events, as well as two anonymous referees and the Editor of Regional Studies for comments and suggestions. All usual caveats apply.

Notes

Since firms without employees at the beginning of the period and/or firms established in the final year were present in the sample, a value of 1 had to be added to SIZE and AGE before the log-transformation in order to prevent a natural logarithm of zero.

A simple comparison of the distribution of the sample and the approximate targeted population by NUTS-4 districts reveals that there is very high correlation between them of 0.99 without logs and 0.94 in logs; respectively, 0.91 and 0.96 without Prague in the sample.

Of course, it would be also preferable to take into account data on research and development activity, but unfortunately this information is not available for the NUTS-4 regions. Nevertheless, the indicator of employment in business services (70–74 codes according to NACE, rev. 1.1) covers firms specialized in providing research and development, technical, engineering, and other consultancy services, which comprised about 30% of total research and development employment in the business sector in 2000, so that agglomeration of this activity is at least partly represented.

A specialized statistical software called Hierarchical Linear and Non-linear Modeling (HLM) version 6.04 was used to estimate the equations. For details on the estimation procedure, see Raudenbush et al. (2004).

As noted above, the estimated coefficients can be transformed from the log-odds back into the expected probability to innovate by using the inverse of the logit link function, which then allows marginal effects to be computed in a standard way.

Since the HLM (version 6.04) package assumes that the variances might not be normally distributed, a Chi-square (χ ²) test of the residuals is performed (Raudenbush et al., 2004). Nevertheless, this should be interpreted with caution because the variances are bounded at zero by definition, while the residuals are generally expected to be non-zero, so that the meaning of their statistical significance is not the same as for an ordinary variable. Luke (2004, p. 32), therefore, recommends to put more weight on interpreting and comparing between estimates magnitude of the random effects rather than their significance.

As already noted above, SIZE and AGE have been standardized before the estimate, FOR and SEC have meaningful zero points by definition, and, therefore, ‘zero scores on the level 1 predictors’ in this specification of the model refer to a typical (average size and age) domestic-owned firm in a non-manufacturing sector located in an otherwise average region (zeros on the regional random effects).

Boschma and Weterings (2005, pp. 576–577) considered using a multilevel model to analyse innovative productivity of software firms in the Netherlands, but they concluded that the differences between regions were not statistically significant enough to justify moving beyond a single-level model. Arguably, they might have been prevented from using multilevel analysis by a very low number of observations per region (169 firms divided into forty regions) in their data set.